# Copyright 2025-present the HuggingFace Inc. team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from __future__ import annotations import warnings from dataclasses import asdict from enum import Enum from typing import Optional import torch import torch.nn as nn from tqdm import tqdm from peft.tuners.tuners_utils import BaseTuner, BaseTunerLayer, check_target_module_exists from peft.utils import ( TRANSFORMERS_MODELS_TO_SHIRA_TARGET_MODULES_MAPPING, ModulesToSaveWrapper, _get_submodules, ) from .config import ShiraConfig from .layer import Linear, ShiraLayer class ShiraModel(BaseTuner): """ Creates a Sparse High Rank Adapter (SHiRA) Model from a pretrained model. Args: model ([`~transformers.PreTrainedModel`]): The model to be adapted. config ([`ShiraConfig`]): The configuration of the SHiRA model. adapter_name (`str`): The name of the adapter, defaults to `"default"`. Returns: `torch.nn.Module`: The SHiRA model. Example: ```py >>> from transformers import AutoModelForCausalLM >>> from peft import ShiraConfig, get_peft_model >>> base_model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m") >>> config = ShiraConfig(r=32) >>> model = get_peft_model(base_model, config) ``` **Attributes**: - **model** ([`~transformers.PreTrainedModel`]) -- The model to be adapted. - **peft_config** ([`ShiraConfig`]): The configuration of the SHiRA model. """ prefix: str = "shira_" def _check_new_adapter_config(self, config: ShiraConfig) -> None: """ A helper method to check the config when a new adapter is being added. Raise a ValueError if there is something wrong with the config or if it conflicts with existing adapters. """ for existing_config in self.peft_config.values(): if existing_config is config: # skip the current config continue @staticmethod def _check_target_module_exists(shira_config, key): return check_target_module_exists(shira_config, key) def _create_and_replace( self, shira_config, adapter_name, target, target_name, parent, current_key, **optional_kwargs, ): if current_key is None: raise ValueError("Current Key shouldn't be `None`") bias = hasattr(target, "bias") and target.bias is not None kwargs = {} kwargs["bias"] = bias if shira_config.mask_type == "random": kwargs["random_seed"] = shira_config.random_seed for k, v in optional_kwargs.items(): kwargs[k] = v if isinstance(target, Linear): mask = ( shira_config.mask_fn(target.base_layer, shira_config.r, **kwargs) if shira_config.mask_fn is not None else None ) target.update_layer( adapter_name, mask, shira_config.r, init_weights=shira_config.init_weights, ) else: new_module = self._create_new_module(shira_config, adapter_name, target, **kwargs) if adapter_name not in self.active_adapter: # adding an additional adapter: it is not automatically trainable new_module.requires_grad_(False) self._replace_module(parent, target_name, new_module, target) @staticmethod def _replace_module(parent, child_name, new_module, child): setattr(parent, child_name, new_module) # It's not necessary to set requires_grad here, as that is handled by # _mark_only_adapters_as_trainable # child layer wraps the original module, unpack it if hasattr(child, "base_layer"): child = child.base_layer if not hasattr(new_module, "base_layer"): new_module.weight = child.weight if hasattr(child, "bias"): new_module.bias = child.bias if getattr(child, "state", None) is not None: if hasattr(new_module, "base_layer"): new_module.base_layer.state = child.state else: new_module.state = child.state new_module.to(child.weight.device) meta = torch.device("meta") # dispatch to correct device for name, module in new_module.named_modules(): if "shira_" in name: if not any(p.device == meta for p in module.parameters()): module.to(child.weight.device) def _mark_only_adapters_as_trainable(self, model: nn.Module) -> None: for n, p in model.named_parameters(): if self.prefix not in n: p.requires_grad = False @staticmethod def _create_new_module(shira_config, adapter_name, target, **kwargs): fan_in_fan_out = shira_config.fan_in_fan_out _ = kwargs.pop("bias", False) if isinstance(target, BaseTunerLayer): target_base_layer = target.get_base_layer() else: target_base_layer = target if isinstance(target_base_layer, torch.nn.Linear): if fan_in_fan_out: warnings.warn( "fan_in_fan_out is set to True but the target module is `torch.nn.Linear`. " "Setting fan_in_fan_out to False." ) fan_in_fan_out = shira_config.fan_in_fan_out = False else: raise ValueError( f"Target module {target} is not supported. Currently, only the following modules are supported: " "`torch.nn.Linear`." ) mask = ( shira_config.mask_fn(target_base_layer, shira_config.r, **kwargs) if shira_config.mask_fn is not None else None ) new_module = Linear( target, mask, adapter_name, shira_config.r, fan_in_fan_out, init_weights=shira_config.init_weights, **kwargs, ) return new_module def __getattr__(self, name: str): """Forward missing attributes to the wrapped module.""" try: return super().__getattr__(name) # defer to nn.Module's logic except AttributeError: if name == "model": # see #1892: prevent infinite recursion if class is not initialized raise return getattr(self.model, name) def get_peft_config_as_dict(self, inference: bool = False): config_dict = {} for key, value in self.peft_config.items(): config = {k: v.value if isinstance(v, Enum) else v for k, v in asdict(value).items()} if inference: config["inference_mode"] = True config_dict[key] = config return config def _set_adapter_layers(self, enabled=True): for module in self.model.modules(): if isinstance(module, (BaseTunerLayer, ModulesToSaveWrapper)): module.enable_adapters(enabled) def enable_adapter_layers(self): self._set_adapter_layers(enabled=True) def disable_adapter_layers(self): self._set_adapter_layers(enabled=False) def set_adapter(self, adapter_name): for module in self.model.modules(): if isinstance(module, ShiraLayer): if module.merged: warnings.warn("Adapter cannot be set when the model is merged. Unmerging the model first.") module.unmerge() module.set_adapter(adapter_name) self.active_adapter = adapter_name @staticmethod def _prepare_adapter_config(peft_config, model_config): if peft_config.target_modules is None: if model_config["model_type"] not in TRANSFORMERS_MODELS_TO_SHIRA_TARGET_MODULES_MAPPING: raise ValueError("Please specify `target_modules` in `peft_config`") peft_config.target_modules = set( TRANSFORMERS_MODELS_TO_SHIRA_TARGET_MODULES_MAPPING[model_config["model_type"]] ) return peft_config def _unload_and_optionally_merge( self, merge=True, progressbar: bool = False, safe_merge: bool = False, adapter_names: Optional[list[str]] = None, ): # we cannot use self.prefix as we want to include non-trainable shira parameters key_list = [key for key, _ in self.model.named_modules() if "shira" not in key] desc = "Unloading " + ("and merging " if merge else "") + "model" for key in tqdm(key_list, disable=not progressbar, desc=desc): try: parent, target, target_name = _get_submodules(self.model, key) except AttributeError: continue if hasattr(target, "base_layer"): if merge: target.merge(safe_merge=safe_merge, adapter_names=adapter_names) self._replace_module(parent, target_name, target.get_base_layer(), target) elif isinstance(target, ModulesToSaveWrapper): # save any additional trainable modules part of `modules_to_save` setattr(parent, target_name, target.modules_to_save[target.active_adapter]) return self.model def delete_adapter(self, adapter_name: str): """ Deletes an existing adapter. Args: adapter_name (str): Name of the adapter to be deleted. """ if adapter_name not in list(self.peft_config.keys()): raise ValueError(f"Adapter {adapter_name} does not exist") del self.peft_config[adapter_name] # we cannot use self.prefix as we want to include non-trainable shira parameters key_list = [key for key, _ in self.model.named_modules() if "shira" not in key] new_adapter = None for key in key_list: _, target, _ = _get_submodules(self.model, key) if isinstance(target, ShiraLayer): target.delete_adapter(adapter_name) if new_adapter is None: new_adapter = target.active_adapter[:] self.active_adapter = new_adapter or [] def merge_and_unload( self, progressbar: bool = False, safe_merge: bool = False, adapter_names: Optional[list[str]] = None ): r""" This method merges the Shira layers into the base model. This is needed if someone wants to use the base model as a standalone model. Args: progressbar (`bool`): whether to show a progressbar indicating the unload and merge process safe_merge (`bool`): whether to activate the safe merging check to check if there is any potential Nan in the adapter weights adapter_names (`list[str]`, *optional*): The list of adapter names that should be merged. If None, all active adapters will be merged. Defaults to `None`. Example: ```py >>> from transformers import AutoModelForCausalLM >>> from peft import ShiraConfig, get_peft_model >>> base_model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m") >>> config = ShiraConfig(r=32) >>> model = get_peft_model(base_model, config) >>> ## [Train the adapter] ## >>> merged_model = model.merge_and_unload() ``` """ return self._unload_and_optionally_merge( progressbar=progressbar, safe_merge=safe_merge, adapter_names=adapter_names ) def unload(self): """ Gets back the base model by removing all the Shira modules without merging. This gives back the original base model. """ return self._unload_and_optionally_merge(merge=False)